(root)/
gcc-13.2.0/
gcc/
testsuite/
gcc.target/
i386/
avx512f-vfmaddXXXsd-2.c
       1  /* { dg-do run } */
       2  /* { dg-options "-O2 -mavx512f" } */
       3  /* { dg-require-effective-target avx512f } */
       4  
       5  #include "avx512f-check.h"
       6  
       7  #include "avx512f-helper.h"
       8  
       9  #define SIZE (128 / 64)
      10  #include "avx512f-mask-type.h"
      11  
      12  static void
      13  calc_add (double *s1, double *s2, double *s3, double* r)
      14  {
      15    r[0] = s1[0] * s2[0] + s3[0];
      16    
      17    int i;
      18    for (i = 1; i < SIZE; i++)
      19      r[i] = s1[i];
      20  }
      21  
      22  static void
      23  calc_add_3 (double *s1, double *s2, double *s3, double* r)
      24  {
      25    r[0] = s2[0] * s3[0] + s1[0];
      26    
      27    int i;
      28    for (i = 1; i < SIZE; i++)
      29      r[i] = s1[i];
      30  }
      31  
      32  void
      33  avx512f_test (void)
      34  {
      35    int i, sign;
      36    union128d res1, res2, res3, res4, res5, res6, res7, src1, src2, src3;
      37    MASK_TYPE mask = MASK_VALUE;
      38    double res_ref1[SIZE], res_ref2[SIZE];
      39  
      40    sign = -1;
      41    for (i = 0; i < SIZE; i++)
      42      {
      43        src1.a[i] = DEFAULT_VALUE;
      44        src2.a[i] = 56.78 * (i + 1) * sign;
      45        src3.a[i] = 90.12 * (i + 2) * sign;
      46        sign = sign * -1;
      47      }
      48    for (i = 0; i < SIZE; i++)
      49      {
      50        res1.a[i] = DEFAULT_VALUE;
      51        res2.a[i] = DEFAULT_VALUE;
      52        res5.a[i] = DEFAULT_VALUE;
      53        res6.a[i] = DEFAULT_VALUE;
      54      }
      55  
      56    calc_add (src1.a, src2.a, src3.a, res_ref1);
      57    calc_add_3(src1.a, src2.a, src3.a, res_ref2);
      58  
      59    res1.x = _mm_mask_fmadd_sd (src1.x, mask, src2.x, src3.x);
      60    res2.x = _mm_mask3_fmadd_sd (src2.x, src3.x, src1.x, mask);
      61    res3.x = _mm_maskz_fmadd_sd (mask, src1.x, src2.x, src3.x);
      62    res4.x = _mm_fmadd_round_sd (src1.x, src2.x, src3.x, _MM_FROUND_NO_EXC);
      63    res5.x = _mm_mask_fmadd_round_sd (src1.x, mask, src2.x, src3.x, _MM_FROUND_NO_EXC);
      64    res6.x = _mm_mask3_fmadd_round_sd (src2.x, src3.x, src1.x, mask, _MM_FROUND_NO_EXC);
      65    res7.x = _mm_maskz_fmadd_round_sd (mask, src1.x, src2.x, src3.x, _MM_FROUND_NO_EXC);
      66  
      67    if (check_union128d (res4, res_ref1))
      68      abort();
      69  
      70    MASK_ZERO (d) (res_ref1, mask, 1);
      71    if (check_union128d (res3, res_ref1))
      72      abort ();
      73  
      74    MASK_ZERO (d) (res_ref1, mask, 1);
      75    if (check_union128d (res7, res_ref1))
      76      abort ();
      77  
      78    MASK_MERGE (d) (res_ref2, mask, 1);
      79    if (check_union128d (res2, res_ref2))
      80      abort ();
      81  
      82    MASK_MERGE (d) (res_ref2, mask, 1);
      83    if (check_union128d (res6, res_ref2))
      84      abort ();
      85  
      86    MASK_MERGE (d) (res_ref1, mask, 1);
      87    if (check_union128d (res1, res_ref1))
      88      abort ();
      89    
      90    MASK_MERGE (d) (res_ref1, mask, 1);
      91    if (check_union128d (res5, res_ref1))
      92      abort ();
      93  }
      94